Applied Mechanics and Materials Vol 772 (2015) pp 268-273
© (2015) Trans Tech Publications, Switzerland
doi:10.4028/www.scientific.net/AMM.772.268
Artificial Neural Network Model for Tool Condition Monitoring in Stone
Drilling
Danko Brezak1,a, Tomislav Staroveski2,b, Ivan Stiperski3,c, Miho Klaic4,d
and Dubravko Majetic5,e
1,2,3,4,5
Faculty of Mechanical Engineering and Naval Architecture, University of Zagreb, Ivana
Lucica 5, Zagreb, HR-.10000, Croatia
a
[email protected], [email protected], [email protected],
d
[email protected], [email protected]
Keywords: Stone drilling, tool wear classification, neural networks
Abstract. This paper explores the possibility of tool wear classification in stone drilling. Wear
model is based on Radial Basis Function Neural Network which links tool wear features extracted
from motor drive current signals and acoustic emission signals with two wear levels – sharp and
worn drill. Signals were measured during stone drilling under different cutting conditions, and then
filtered before tool wear features extraction. Features were obtained from time and frequency
domain. They have been analyzed individually and in combinations. The results indicate tool wear
monitoring capacity of the proposed model in stone drilling, and its potential for simple and costeffective integration with CNC machine tools.
Introduction
Tool wear monitoring is one of the most important segments in the development of fully automated
and highly autonomous CNC machine tools. Except in machine tool diagnostics, it is also necessary
in the implementation of machining process control systems which could prevent tool breakage
and/or maintain predefined tool wear dynamic [1, 2]. Tool wear monitoring in drilling has been
continuously in the research focus for the past 20 years. A number of machine learning algorithms,
sensor combinations and tool wear features have been analyzed and proposed, mainly using metal
and composite materials [3]. Only a few studies considered wear monitoring in stone machining.
They have usually included wear identification of diamond tools applied in cutting and/or milling
using cutting forces sensors [4-7].
The aim of this study was to analyze capabilities of neural network-based tool wear classification
model in stone drilling using cost-effective combination of internal drive signals or currents (instead
of cutting forces) and acoustic emission sensor. For this purpose, a type of Radial Basis Function
Neural Network (RBF NN) algorithm for solving classification types of problems has been chosen.
This type of neural network is known for its learning in one step and a capability of simple and
quick hidden layer structure adaptation. Experiments were conducted using a custom-made machine
tool testbed with open architecture control platform.
Experimental Work
Machine Tool Testbed. Experimental work has been performed using the three-axis bench-top
CNC mini milling machine with an internal and external measurement systems (Fig. 1). The
machine has been retrofitted with the 0.4 kW (1.27 Nm) permanent magnet synchronous motors
with integrated incremental encoders (Mecapion SB04A), corresponding motor controllers
(DPCANIE-030A400 and DPCANIE-060A400), ball screw assemblies, and LinuxCNC open
architecture control (OAC) system [8]. Considering the nature of the drilling process, two types of
signals were sampled from those controllers: vertical or Z-axis feed drive current (IZ) and main
spindle current (IMS). Beside motor drive currents, acoustic emission signals (AE) were also
measured using 8152B piezoelectric AE sensor and 5125 coupler (Kistler) connected to PCIAll rights reserved. No part of contents of this paper may be reproduced or transmitted in any form or by any means without the written permission of Trans
Tech Publications, www.ttp.net. (ID: 89.164.241.35-29/04/15,21:32:07)
Applied Mechanics and Materials Vol. 772
269
DAS4020/12 data acquisition board. Customized measurement software developed in the LabView
environment was used for cutting conditions setup, storage of the acquired signals and NC drilling
cycle program generation. Each measurement started by issuing the trigger signal from the CNC
control system to the measurement systems. Direct observations of the drill cutting edges were
made using the industrial camera type DMK41AF02 equipped with the telecentric lenses type
TC2309. Twist drill type BOSCH CYL-9 (5 mm in diameter) was used to drill 10 mm deep holes in
Adria Grigio Machiato stone samples.
Fig. 1. Experimental setup - work area and sensor placement. (1) X axis feed drive; (2) Y axis feed
drive; (3) Z axis feed drive; (4) main spindle drive; (5) workpiece fixture; (6) three axis force
sensor; (7) test drill; (8) stone workpiece; (9) AE sensor
Data Acquisition and Signal Processing. Motor drive current signals were sampled continuously
at 1 kHz, and AE signals at 2 MHz (0.1s sample duration per hole). Signals were measured with
sharp (SD) and worn drill (WD) using 9 combinations of cutting speeds (10; 30; 50 m/min), and
feed rates (0.05; 0.1; 0.15 mm/rev), which were chosen according to the tool manufacturer
recommendations for this type of drill. Those cutting speeds correspond to spindle speeds of 636.6
rpm, 1909.8 rpm and 3183.09 rpm, respectively. For each combination of machining parameters
experiment was randomly repeated 10 times. Measurements are first taken while drilling with
completely sharp drill and then repeated using the worn drill. After completing the measurements
with the sharp drill, it was then used to drill a number of cycles until it completely worn out. Flank
wear and cracks were observed as a dominant wear features (Fig. 2).
Fig. 2. Images of cutting edges after drilling with sharp (SD) and worn drill (WD) with observable
(arrows pointing to) flank wear area
270
Advanced Research in Aerospace, Robotics, Manufacturing Systems,
Mechanical Engineering and Bioengineering
Before extracting tool wear features, both types of signals were filtered. In the case of current
signals, Butterworth low-pass filter with 2 Hz cut-off frequency was applied. This frequency was
chosen after spectral analyzes of signals using Fast Fourier Transform (FFT). AE signals were
filtered with the Butterworth band-pass filter (frequency range 40-500 kHz), which was in
accordance with the specified frequency or measurement range of the utilized sensor.
Tool Wear Features. After filtration, 8 features were extracted from current signals. First two
features were maximum values of both types of signals (Max_IZ, Max_IMS). They were calculated
based on an average value of the 10% of the highest current values, thus neutralizing eventual
occurrence of transient spikes. The next two features (Area_IZ, Area_IMS) were areas under current
curves related to the machining time. This type of features is closely related to the total amount of
electric energy used in the cutting process. The remaining four features were from the frequency
domain: power of spectral components related to the rotation frequency - RF (P_RF_IZ, P_RF_IMS)
and cutting edges frequency - CF (P_CEF_IZ, P_CEF_IMS) [9]. Since drill has two cutting edges,
CEF was twice as high as RF. Those features were obtained using the FFT algorithm.
Features from AE signals were all extracted from the frequency domain. Measured frequency
range (40-500 kHz) was divided into 7 frequency ranges (50-100 kHz; 100-150 kHz; …; 350-400
kHz), and energy of every range has been taken as a drill wear feature [10]
fh
∫
ψ 2 = Sy df ,
(1)
fl
where Sy is the one-sided PSD function of the AE signal, while fl and fh are lower and upper
frequency values chosen to reflect the energy in the range of interest. Altogether, 15 features have
been extracted from both types of signals (Table 1).
Table 1. List of Drill Wear Features
Acoustic
Emission
Current
Type of
signal
Feature
Description
(I1)
Max_IZ
(I2)
Max_IMS
(I3)
(I4)
(I5)
(I6)
(I7)
(I8)
(AE1)
(AE2)
(AE3)
(AE4)
(AE5)
(AE6)
(AE7)
Area_IZ
Area_IMS
P_RF_IZ
P_RF_IMS
P_CEF_IZ
P_CEF_IMS
ψ2 (AE50-100)
ψ2 (AE100-150)
ψ2 (AE150-200)
ψ2 (AE200-250)
ψ2 (AE250-300)
ψ2 (AE300-350)
ψ2 (AE350-400)
Average of a group of 10% highest absolute IZ current values
obtained from feed motor drive (vertical or Z-axis)
Average of a group of 10% highest IMS current values obtained
from main spindle motor drive
Area under the IZ=f(time) curve
Area under the IMS=f(time) curve
Power of rotational frequency component of the IZ signal
Power of rotational frequency component of the IMS signal
Power of cutting edges frequency component of the IZ signal
Power of cutting edges frequency component of the IMS signal
Energy of the AE signal in the frequency range 50-100 [kHz]
Energy of the AE signal in the frequency range 100-150 [kHz]
Energy of the AE signal in the frequency range 150-200 [kHz]
Energy of the AE signal in the frequency range 200-250 [kHz]
Energy of the AE signal in the frequency range 250-300 [kHz]
Energy of the AE signal in the frequency range 300-350 [kHz]
Energy of the AE signal in the frequency range 350-400 [kHz]
Applied Mechanics and Materials Vol. 772
271
RBF Neural Network
Chosen NN algorithm is based upon a well-known feedforward three-layered RBF NN architecture,
where the matrix/vector of synaptic weights c is calculated in the learning phase using the
expression
c = H+ y ,
(2)
where y stands for the matrix/vector of desired output values and H+ is Moore – Penrose
pseudoinverse of the matrix of hidden layer neuron RBF outputs or activation function outputs (H).
The pseudoinverse is defined as follows
−1
H+ = (HTH ) HT .
(3)
In the testing phase, the matrix of desired output values y is obtained from the expression
y = Hc .
(4)
Elements of matrix H are determined according to the expression
H ij = e
1
− r ij2
2
, i = 1, ..., N, j = 1, ..., K
(5)
where rij is the Mahalanobis distance between vector composed from ith element of all input vectors
(tool wear features) and jth hidden layer neuron (or hidden layer neuron center). Squared
Mahalanobis distance is calculated using the expression
T
rij2 = ( x i − t j ) Σ j−1 ( x i − t j ) ,
(6)
where Σj is a covariance matrix belonging to the group of learning samples that are connected to the
jth hidden layer neuron, xi is the L-dimensional vector composed from ith element of all L input
vectors and tj is L-dimensional vector of the jth hidden layer neuron center. Covariance matrix is
quadratic matrix with non-zero elements (squared σ vector components) on main diagonal and
zeros elsewhere,
σ12 0
0 


Σj =  0 0  .
 0 0 σL2 


(7)
Since every center is defined in the learning phase based on the group of network input elements,
vector σ is composed from the maximal Euclidian distances between learning samples belonging to
the analyzed group and the center of that group, regarding to all (L) dimensions separately,
σ g j = max
{
z pg − t g , p = 1,..., LK G
}  , g = 1,..., L ,
j
(8)
where zpg is the gth component of the pth sample of the jth group which is defined with LKG
numbers of samples, and tg is gth component of the jth group center vector (jth hidden layer neuron
center vector).
Hidden layer neuron centers are defined using a method which helps teacher to quickly
determine network structure regarding to the nature of the learning problem and desirable network
272
Advanced Research in Aerospace, Robotics, Manufacturing Systems,
Mechanical Engineering and Bioengineering
generalization characteristics. Grouping of tool wear features or network input elements and centers
calculations are based on the parameter βC. Higher βC reduces the number of hidden layer neurons
and vice versa (for βC=0 every input vector forms one hidden layer neuron or its center). Hidden
layer configuration method is in detail explained in [11].
Results
With 9 combinations of machining parameters, 10 measurements for each combination, and 2 drill
wear levels or classification groups, 180 sets of samples were collected in total. Five out of 10
samples of repetitive measurements for each combination of machining parameters were used in the
learning phase, and the remaining five participated in the formation of data sets used in the testing
phase of the RBF NN classifier.
In order to analyze capacity of chosen features for drill wear classification, and to find
combination(s) which provide the best classification performance, learning/testing procedure was
divided into several steps. In the first step, every feature has been analyzed separately using full
hidden layer structure (βC=0). Based on these first results, further analyzes of different feature
combinations have been performed (also with βC=0). Feature combinations have expectedly
achieved higher classification accuracy than the individual features, and the results of chosen
combinations are presented in Table 2. All presented results were achieved using cutting speed and
feed rate as two additional NN inputs, and classification success rate in the learning phase was
100%.
Table 2. Classification Results – Accurately Classified Samples, (%)
TEST
RBF NN
Feature
structure
T1
T2
T3
T4
T5
Avg.
I1+I2+I3+I4
88.9 94.4 88.9 94.4 100
6-90-2
93.3
I5+I6+I7+I8
66.7 83.3 83.3 66.7 83.3
6-90-2
76.7
10-90-2
83.3 88.9 94.4 72.2 94.4
86.6
I1+I2+...+I8=ΣI
AE4+AE5
88.9 77.8 83.3 72.2 88.9
4-90-2
82.2
9-90-2
74.5
AE1+AE2+…+AE7=ΣΑΕ 77.8 77.8 77.8 77.8 61.1
I1+I2+I3+I4+AE4+AE5
83.3 94.4 94.4 100 94.4
8-90-2
93.3
17-90-2
83.3
88.9
94.4
72.2
83.3
84.4
ΣI+ ΣAE
I1+I2+I3+I4+AE4+AE5
88.9 94.4 88.9 94.4 88.9
91.1
8-58-2 (βC≠0)
Combination of time domain features extracted from IZ and IMS current signals has shown the
highest average classification accuracy based on 5 analyzed tests (93.3%). The same result was
achieved when this combination was further extended by two extra features from AE signals (AE4,
AE5). These two features accomplished the best individual classification accuracy among the
features from AE signals. Other combinations did not manage to reduce average classification error
below 10%, but their results are still more than acceptable.
At the end, the best combination of features extracted from both types of signals was analyzed
once more with the reduced number of hidden layer neurons (βC>0) to find out the RBF NN
structure still capable to provide satisfactory generalization characteristics. It can be noticed that
RBF NN with the number of hidden layer neurons decreased by 30% (58 vs. 90) managed to
maintain classification accuracy higher than 90%.
Summary
In this work a type of Radial Basis Function Neural Network algorithm has been applied for tool
wear classification under different cutting conditions in stone drilling. Tool wear features were
obtained from Z-axis (vertical axis) feed drive current, main spindle drive current, and acoustic
Applied Mechanics and Materials Vol. 772
273
emission signals. They were then used for classification of drill wear into two groups: sharp and
completely worn drill. Every feature was analyzed separately and in combinations.
The accomplished results suggest potential of the analyzed model for tool wear monitoring
during stone drilling. Most successful model outputs were achieved using only features from
servomotor current signals, thus supporting the idea of reliable and cost-effective monitoring
system without the usage of force sensor. Practically identical result has been achieved using the
combination of selected features from current and acoustic emission signals. This is also very
important, since it is a wide known fact that tool wear is highly non-linear and partially stochastic
process which cannot be reliably identified and monitored in the real industrial environment using
only one type of signal. Furthermore, AE sensors are low-cost and easily integrated into the
machine tool structure.
The proposed tool wear model is very simple to implement in the CNC control system and
shows the potential for tool wear monitoring in drilling. It was analyzed with experimental data
obtained while drilling a single type of stone material with a single type of twist drill. Future
experimental work will therefore be extended to different types of stone materials, drill diameters
and geometries, as well as more than two drill wear levels.
References
[1] S. Liang, R.L. Hecker, and R.G. Landers: ASME Journal of Manufacturing Science and
Engineering, Vol. 126 (2004), pp. 297–310.
[2] R. Teti, K. Jemielniak, G. O’Donnell, and D. Dornfeld: CIRP Annals - Manufacturing
Technology, Vol. 59 (2010), pp. 717-739.
[3] E. Jantunen: Journal of Machine Tools and Manufacture, Vol. 42 (2002), pp. 997-1010.
[4] W. Polini and S. Turchetta: The International Journal of Advanced Manufacturing Technology,
Vol. 35 (2007), pp 454-467.
[5] W. Polini and S. Turchetta: Advances in Mechanical Engineering, Vol. 2009 (2009).
[6] S. Turchetta: The International Journal of Advanced Manufacturing Technology, Vol. 61
(2012), pp. 441-448.
[7] J. Kenda and J. Kopac: Strojniški vestnik - Journal of Mechanical Engineering, Vol. 55 (2009),
pp. 775-780.
[8] T. Staroveski, D. Brezak, T. Udiljak, and D. Majetic: Annals of DAAAM International 2011,
Vol. 22 (2011), pp. 0023-0024.
[9] P.W. Prickett, C. Johns: International Journal of Machine Tools & Manufacture, Vol. 1 (1999),
pp.105-122.
[10] C. Scheffer, P.S. Heyns, and F. Klocke: International Journal of Machine Tools &
Manufacture, Vol. 43 (2003), pp. 973–985.
[11] D. Brezak, T. Udiljak, K. Mihoci, D. Majetic, B. Novakovic, and J. Kasac: Proceedings of
International Joint Conference on Neural Networks - IJCNN (2004), pp. 1859-1863.